Social Media Sensors for Weather-Caused Outage Prediction Based on Spatio–Temporal Multiplex Network Representation

This study investigates severe weather events that lead to power outages. Despite extensive research on using social media during disasters, little work has focused on combining social media information with power outage data. To address this limitation, we propose a novel and effective approach to enhance the prediction accuracy of weather-related power outages by learning a spatio-temporal multiplex network that integrates information on the impact of inclement weather on the residents extracted from their social media posts with relevant weather, geographic, and grid topology data. In the multiplex network framework, the outage risk is estimated using logistic regression, neural network, support vector machine, random forest, xgboost, and decision tree classifiers through which machine learning models are applied separately on individual layers and jointly on multiplex networks representing layers capturing information related to weather, lightning, and vegetation. Experiments were conducted to predict the risk for weather-related power outages three hours in advance at a county level in five states of the Pacific Northwest region from November 2021 to April 2022. Evidence suggests that using vegetation information improves the quality of all models compared to relying on weather and lighting layers alone. Integrating an additional layer on the impact of inclement weather on the community residents retrieved from public social media posts (Twitter, Reddit) with weather, lightning, and vegetation layers improves the accuracy of outage prediction by $5\%-7\%$ . The results demonstrate that the proposed spatio-temporal multiplex network-based approach offers beneficial insights for predicting power outages three hours ahead at the county level.